<![CDATA[Amazon Reviews'23 Gift Cards tests whether review text helps when users are sparse and products popularity-dominated. Its 152,410 reviews cover 132,732 users, 1,137 parent products, and complete metadata. Collapsing repeat user-product pairs yielded 125,704 verified positives rated at least four stars. Leave-two-out retained 120,154 training interactions and evaluated 2,769 active users over all 1,009 items with one frozen pre-validation history. Ten recommenders covered popularity, transitions, item/user retrieval, low-rank and corrupted-view reconstruction, BPR, graph propagation, review retrieval, and gated fusion. Separate branches tested first-observation cold start, clipped-noise client learning, and post-hoc evidence attribution. Markov achieved NDCG@10 0.2346 and HR@10 0.4056; gating gave it all weight, so reviews neither improved warm ranking nor caused recommendations. In the 2021 cold-start proxy, metadata TF-IDF reached micro NDCG@10 0.4613 over 113 candidates, but item-macro and dominant-target-excluded NDCG@10 scores were 0.0627 and 0.0605. Across 500 cases, the cited template reached 0.998 lexical support and 1.000 source localization; shuffling preserved support but reduced user-profile cosine from 0.1683 to 0.0392. Thus ranking, support, alignment, and causal faithfulness diverge.]]>
Copyrights © 2026
